Key takeaways
- Most agency clients need 50 to 120 tracked prompts. Under 20 is a sanity check, not a measurement, and above 500 most teams stop reading the data.
- Client type changes the number: a local service business can live on 30 to 50, a multi-market e-commerce brand may need 250 or more.
- Organic and comparison prompts produce crowded, noisy answers, so they deserve more of your budget than branded prompts.
- Build one template, share it across accounts, and keep a fixed core set so you can tell engine changes from real change.
- Start with 40 to 60 prompts, act on what you see, and only then expand.
The short answer
There is no universal number, but there is a useful range. Refine's sizing guide puts agencies tracking one client at 50 to 120 prompts, sharing a template across accounts. Chatter Buzz Media suggests starting with 40 to 60 prompts across five to ten core topics and growing only where answers are unstable. SE Ranking's guide to choosing prompts has a similar flavor: a small awareness slice (10 to 20 prompts) inside a larger set.
I think those ranges are about right, and I'd add one agency-specific opinion: the real constraint is not statistical, it's attention. You have maybe a dozen clients. If each has 400 prompts, nobody is reading the report. Pick a size you can actually discuss on a monthly call.

Why prompt count changes your numbers
AI answers are probabilistic. Ask the same question twice and you can get different brands, in a different order, with different sources. Your visibility score is an estimate, and estimates tighten as the sample grows.
The arithmetic is simple. Track 10 prompts and a single answer flipping moves your score by 10 points. Track 100 and the same flip moves it by 1 point. So a client score that jumps from 40% to 52% on a tiny set is usually variance, not a win you should send a celebratory email about.
Why organic prompts need more volume
Promptwatch Data's report on how many unique domains ChatGPT cites per response (Sep 28 to Oct 4, 2026) shows how crowded these answers are. Across responses with at least one citation, 47.6% cite 10 or more distinct domains. Split by prompt type:
| Prompt type | Cite 10+ domains | Cite 3 or fewer | Most common count |
|---|---|---|---|
| Organic (no brand named) | 52.5% | 4.5% | n/a |
| Competitor comparison | 49.1% | under 1% single-domain | n/a |
| Brand-specific | 21.3% | 13.8% | 5 domains (14.9%) |
Branded answers are tight and predictable. Organic and comparison answers draw from a wide pool of sources, so any single prompt tells you little. That is the sizing argument in one table: spend your prompts where the variance is.
One caveat. Promptwatch's older average sources per response report (dated Sep 2025) cites roughly 5 sources per ChatGPT web-search answer, which conflicts with the newer 10-plus finding. Treat the 5 figure as dated and use the October 2026 numbers.
Run each prompt more than once
Prompt count is only half of sample size. Refine recommends running each prompt several times per engine and reporting visibility as a rate. Whichever tool you use, check that it repeats runs rather than showing a one-off yes or no.
Sizing by client type
Here is how I'd size the common agency client types. These are starting points built from the ranges above, not laws.
| Client type | Starting prompts | Ceiling before you should split projects | What drives the count |
|---|---|---|---|
| Local service business (dentist, plumber, law firm) | 30 to 50 | 80 | Services times a handful of locations |
| Focused B2B SaaS, one product | 50 to 100 | 150 | Use cases, alternatives, integrations |
| Multi-product SaaS or several segments | 100 to 250 | 250 | Product lines and personas |
| E-commerce, small catalog | 60 to 120 | 200 | Category and product comparison prompts |
| E-commerce or marketplace, large catalog | 150 to 500 | 500 | Categories, brands, shopping intent |
| Multi-country or multi-language brand | 100 to 250 per market | 500 total per monitor group | Each market is its own set |
| Multi-location franchise | 40 to 60 per location type, plus a shared core | 300 | Location-specific "near me" prompts |
| Enterprise brand with several divisions | 250 to 500 | split by division | Divisions with different buyers |
Local and multi-location clients
Prompts here are repetitive by design: service plus city plus a modifier. Resist writing 200 near-duplicates. Thirty to fifty well-chosen prompts, then pin the monitor to the actual location. Tools with location-level tracking, such as Promptwatch's local GEO targeting at exact coordinates or Birdeye Search AI, matter more than raw prompt count here.

B2B SaaS
The sweet spot is 50 to 100 for a single product. Weight the set toward alternatives, comparisons and use-case prompts, because that is where shortlists form. Add integration and pricing prompts for the decision stage.
E-commerce
Catalog size drives everything. Group products into categories and track category-level and comparison prompts first, then product-level prompts only for hero items. Product pages are the most cited content type in ChatGPT at 22.4% of citations (Promptwatch Data, Sep 28 to Oct 4, 2026), so product-level visibility is worth measuring. For catalog-specific tracking, look at ChatGPT Shopping insights in Promptwatch or dedicated tools like Ranketta and Alhena AI.
Regulated and vertical clients
Healthcare, travel, software and automotive each have their own citation patterns. Promptwatch Data publishes vertical reports, for example software prompts where GitHub, Microsoft's docs and G2 lead, and travel prompts where booking and search sites earn more than half of citations. Use them to decide which source types your prompts should be probing.
How to allocate prompts across intent
Once you have a number, split it. Refine's suggested allocation looks like this for the later stages:
- Consideration, about 35%: comparisons, alternatives and shortlists.
- Decision, about 25%: pricing, integrations, fit for a specific profile.
- Brand and reputation, about 15%: what AI says when someone asks about the client directly.
That leaves roughly a quarter for awareness and problem-stage prompts, which fits SE Ranking's suggestion of 10 to 20 awareness prompts in a mid-sized set. For a 100-prompt client, that is about 25 awareness, 35 consideration, 25 decision and 15 brand prompts.
Two habits help. Tag every prompt with its intent and topic so you can report by segment. And add personas where buyers genuinely ask differently; a CFO and a developer rarely phrase the same question the same way.
Build one template, then customize
The agency advantage is repetition. Build a template once, then adapt about 60 to 70% of it per client.
- Start with 15 to 20 core service-line or topic slots (Rocket's agency guidance for shared prompt libraries uses the same 15 to 20 starting range, though that's for working prompts rather than tracked ones).
- Fill each slot with an awareness, consideration and decision variant.
- Add the client's brand prompts and two or three competitor comparison prompts per main rival.
- Pull candidate prompts from Search Console queries, People Also Ask, sales call transcripts and support tickets rather than guessing.
- Review after the first two weeks and cut prompts that return identical, uninformative answers.
Query fan-outs help here too. A single prompt can trigger several retrieval searches, and Promptwatch's ChatGPT query fanout data shows that behavior shifts over time (average fan-outs per response were 2.15 in December 2025 and about 1.0 by April 2026). Looking at the actual fan-out queries tells you which sub-questions deserve their own prompts.
Keep a fixed core, rotate the edges
Engines change under you. ChatGPT's average citations per response dropped around the GPT-5.3 rollout on March 4, 2026, and reddit.com's share of ChatGPT citations fell from about 4% to 0.5% on August 14, 2026. If you rewrite your prompt set every month, you can never tell whether the client improved or the engine moved.
My rule: lock about 70% of the set as a core you never touch within a quarter, and use the remaining 30% for new questions, seasonal topics and experiments. Review the core quarterly.
Common sizing mistakes
- Tracking only branded prompts. The score looks great and means nothing, since the client was going to be mentioned when asked about by name.
- Chasing 1,000 prompts before acting on the first 50. The data gets ignored and decisions slow down.
- Tracking too few. At 10 or 15 prompts, one answer changing swings the headline.
- Duplicating near-identical prompts to pad volume. Ten rewordings of one question is one data point with extra cost.
- Mixing markets in one set. Language and location should be fixed per monitor so Dutch and US results never blend.
- Ignoring prompt type. Per the domains report above, organic prompts are where answers are most crowded, so underweighting them hides the real competition.
Budgeting prompts across a client roster
Plans are priced by prompts or responses, so sizing is also a margin question. Rough math: 10 clients at 80 prompts is 800 prompts, and if each runs daily across several engines, responses add up fast. Check whether a tool counts prompts, responses, or both.
A few tools worth comparing for agency use:
| Tool | Agency angle |
|---|---|
| Promptwatch | Agency plans (Kick-off $199/mo, Growth $399/mo, Scale $799/mo) with unlimited projects, prompts and seats; limits are on responses instead |
| Search Party | Built specifically around agency workflows |
| Scrunch AI | Monitoring for brands and agencies |
| Peec AI | Simple tracking with prompt suggestions |
| Otterly.AI | Budget-friendly monitoring |
The pricing model matters more than the sticker. When a plan caps prompts, you ration clients. When it caps responses, you ration frequency, which is usually the easier tradeoff to manage. Promptwatch is the one I'd look at first for agency work because its agency plans don't cap prompts or seats, and it goes past tracking into crawler logs, content gap analysis and reporting. Confirm current pricing at promptwatch.com/pricing.



For more options, the GEO software directory at bestgeosoftware.com lists many more platforms, and the Promptwatch agency page covers the agency-specific setup.
A simple rollout plan
Weeks 1 and 2
Track 40 to 60 prompts per client, split across intent. Run them daily. Don't report scores yet; just check that answers make sense and the brand and competitor detection works.
Weeks 3 to 6
Prune dead prompts, add 20 to 40 where answers vary a lot, and lock the core set. Produce a first baseline report with visibility, share of voice and the top cited sources.
Month 2 onward
Report trends month over month on the fixed core. Use the rotating 30% for client-specific experiments. Expand toward the next tier in the sizing table only if the client has genuinely distinct segments or markets.
FAQ
Is 50 prompts enough for a small client?
For a local or single-product business, yes. Fifty well-chosen prompts, run repeatedly, gives a trustworthy trend.
Should I track the same prompts for every client?
Share the structure, not the wording. Keep the intent split and topic slots the same, and write client-specific prompts inside them.
How often should I change the set?
Review quarterly. Within a quarter, change only the rotating edge so your trend lines stay comparable.
When do I need more than 250 prompts?
When the client has several markets, divisions, or a large catalog. Even then, consider splitting into separate monitors so each report stays readable.
The bottom line
Start at 50 to 60 prompts per client, weight toward organic and comparison queries, lock a core, and scale only when a client's structure demands it. If a report has more prompts than anyone discusses on the monthly call, it's too big.

